← LeaderboardThe Skill Leaderboard

model-onboard

Implement an upstream pretrained model natively in BioIR using its optimized components, with checkpoint conversion and numerical validation. Use when adding a new model or architecture to BioIR; includes data pipeline integration when requested.

Published by NVIDIA ↗NVIDIA/skills

Try your first task

A source-grounded starter task has not been verified for this skill yet.

SKILL.md

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.# SPDX-License-Identifier: Apache-2.0## Licensed under the Apache License, Version 2.0 (the "License");# you may not use this file except in compliance with the License.# You may obtain a copy of the License at## http://www.apache.org/licenses/LICENSE-2.0## Unless required by applicable law or agreed to in writing, software# distributed under the License is distributed on an "AS IS" BASIS,# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.# See the License for the specific language governing permissions and# limitations under the License.name: model-onboarddescription: Implement an upstream pretrained model natively in BioIR using its optimized components, with checkpoint conversion and numerical validation. Use when adding a new model or architecture to BioIR; includes data pipeline integration when requested.license: Apache-2.0metadata:   author: Tien Pham

Model Onboarding

Input: upstream source and revision, a pretrained checkpoint, a writable BioIR checkout, and the requested model features. Output: a native BioIR implementation, checkpoint conversion, tests and validation evidence, and data pipeline integration when in scope.

When the request is an offline decision about a recorded input, write that decision and stop. Do not clone, implement, or run a model.

The procedure lives in the BioIR checkout, written $BIOIR below. Read $BIOIR/docs/advanced/port-model.md in full and follow it. Without a checkout, clone https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime and work on a branch. The guides and the bench-perf-oss skill named below are in that checkout, so stop and report if you can reach neither a checkout nor the clone. Reuse BioIR's modules where their math matches upstream, and implement the architecture's differences. An adapter that only swaps one module into the upstream model does not complete a native model port; for that task, follow $BIOIR/docs/advanced/accelerate-custom-model.md instead.

Establish the Scope

Use the source, checkpoint, checkout, and working directory already supplied in the conversation. Ask only for missing information that affects the port: the authoritative upstream revision and checkpoint, the model variant, and whether full data pipeline integration is required.

Work on a branch in the BioIR checkout. Keep upstream read-only and its dependencies in a separate environment. Store captures, reference outputs, validation scripts, and reports in an external working directory; keep reusable implementation and tests in BioIR.

Before implementing, write a feasibility report in the working directory. Map the upstream components to BioIR components, explain mathematical and layout differences, list missing operations and input types, and state the implementation and validation scope. Resolve an ambiguous architecture or checkpoint with the user; routine implementation choices can proceed.

Execution and Evidence

  • Inspect the available hardware before scheduling validation. Without a supported GPU, continue source analysis, implementation, and checks that run on CPU. Save the GPU validation commands and mark those checks pending; never report the port as validated from static checks alone.
  • Load checkpoints with weights_only=True. Account for checkpoint parameters and buffers: fail on a tensor that is unread or read twice, and explain intentionally unused tensors and shared weights. Never hide missing mappings with non-strict loading.
  • Establish module equivalence on identical captured inputs before relying on end-to-end quality scores, then compare the whole model with the upstream's spread between repeated runs. Record upstream and BioIR revisions, checkpoint identity, settings, tolerances, commands, and artifacts.
  • When full pipeline support is requested, follow $BIOIR/docs/advanced/port-data-pipeline.md and $BIOIR/docs/advanced/validate-data-pipeline.md. Reuse an existing pipeline only after checking its feature and output semantics against upstream.
  • Benchmark only after correctness checks pass. For supported folding models, use the bench-perf-oss skill at $BIOIR/skills/bench-perf-oss/SKILL.md; retain its scope and applicability rules.

Report

State what was implemented, which BioIR components were reused or extended, how checkpoint tensors were accounted for, and how to construct and run the model. List every check as PASS, FAIL, or PENDING with the command, artifact path, and sample scope. Give the exit code for executed checks and the reason for pending checks. Include documented differences, unsupported features, pending GPU work, and the working directory.